Questions on result randomness

[qiyan98]

1. Hyper-parameter auto-selection randomness.
   According to P11 of your paper, in each evaluation, two coefficients ε and ε_s are first automatically selected and then used in the Langevin dynamics sampling to compute the MMD results. However, when running the experiments, we found the selected coefficients may change each time (non-deterministic). This may impact the final MMD, and we are unsure if it is expected. If possible, would you mind sharing the hyper-parameters that you used to obtain your results?

2. MMD performance difference.
   Besides, we ran the code at your repo, did training from scratch, and reproduced some MMD results different than those in your paper. In our reproduction, the MMD is worse than yours most of the time. As our hardware and software environment is different than what you used, the difference in final performance may be caused by some other issues/bugs unknown to us yet. We would appreciate it if you could offer some suggestions.

Many thanks,
Qi